A Bayesian Approach to inferring vascular tree structure from 2D imagery

更新时间:2023-05-28 17:05:01 阅读量: 实用文档 文档下载

说明:文章内容仅供预览,部分内容可能不全。下载后的文档,内容与下面显示的完全一致。下载之前请确认下面内容是否您想要的,是否完整无缺。

We describe a method for inferring tree-like vascular structures from 2D imagery. A Markov Chain Monte Carlo (MCMC) algorithm is employed to produce approximate samples from the posterior distribution given local feature estimates, derived from likelihood

ABAYESIANAPPROACHTOINFERRINGVASCULARTREESTRUCTUREFROM

2DIMAGERY

ElkeTh¨onnes,AbhirBhalerao,WilfridKendall,andRolandWilson

DepartmentsofComputerScienceandStatistics

UniversityofWarwick,UK

elke|wsk@stats.warwick.ac.uk

abhir|rgw@dcs.warwick.ac.uk

ABSTRACT

Wedescribeamethodforinferringtree-likevascularstruc-turesfrom2Dimagery.AMarkovChainMonteCarlo(MCMC)algorithmisemployedtoproduceapproximatesamplesfromtheposteriordistributiongivenlocalfeatureestimates,derivedfromlikelihoodmaximisationforaGaus-sianintensitypro le.Amultiresolutionscheme,inwhichcoarsescaleestimatesareusedtoinitialisethealgorithmfor nerscales,hasbeenimplementedandusedtomodelreti-nalimages.Resultsarepresentedtoshowtheeffectivenessofthemethod.

1.INTRODUCTION

Theproblemofinferringvascularstructurefromimagedataisanimportantone,especiallyintheareaofsurgicalplan-ning,whichrequiresbothef cientcomputationandeffec-tiveuseofpriorknowledge.Previousworkintheareahastendedtofocusonthemodellingofspeci cvascularfea-tures[1]ortouseapproachessuchasadaptivethresholding[4].

Theaimoftheworkdescribedhereistoformulateageneralmethodfortheinference,whichcanbeappliedintwoorthreedimensionsandmakeseffectiveuseofpriorknowledge,yetwhichissuf cientlygeneraltobeappliedtoawiderangeofproblems.Thecommonstatisticalmeth-odsforsuchmedicalimageanalysishavetypicallyusedlikelihoodtechniques,suchasExpectation-Maximisation(EM)[6,5].AlthoughEMmethodscanbeef cientcom-putationally,theyhaveonlylimitedscopeforincorporatingpriorknowledge.AmorepowerfulwayofincludingpriorinformationistouseaBayesianmethod,suchasmaximumaposteriori(MAP)estimation.Theprincipaldif cultywithBayesiantechniquesisacomputationalone:theynormallyrequiretheuseofMarkovchainMonteCarlo(MCMC)al-gorithms,whichmayrunforhundredsofthousandsofiter-ationstoyieldreliableresults[3].Thishasrestrictedtheir

We describe a method for inferring tree-like vascular structures from 2D imagery. A Markov Chain Monte Carlo (MCMC) algorithm is employed to produce approximate samples from the posterior distribution given local feature estimates, derived from likelihood

displacementtobeasmalllinearmultipleofthedisplace-mentoftheparentvertexfromthegrandparentvertex(an“AR(1)”tree)(righthandof gure1).Tomodelintensity,witheachvertexinthetreeweassociateaGaussianker-nelthatrepresentsthespatialgreylevelpro leofthecorre-spondingvesselsegment.

ofTheposteriordistributionforarandomnumber

treesisgivenby

(1)

whereistheimage.Thedistributionof,,penalisesthenumberoftrees;intheexamples,aPoissondistributionwasused.Thisensuresthata‘minimal’expla-nationofthedataisfound.aretheprobabilitiesde ningthedegreeofthebranchingprocessandisthedegreeofthevertex.isthedistribu-tionoftheparametersofthevertex,.Intheexamples,thisisanautoregressive(AR(1))process.Theparametersrepresentthepositionsofvertices,,andtheampli-tudeandwidthparametersoftheedges,whileisthelikelihoodfunction.TheobservationmodelisbasedontheapproximationoflinearstructuresbyasumofGaussianker-nels:

parent

We describe a method for inferring tree-like vascular structures from 2D imagery. A Markov Chain Monte Carlo (MCMC) algorithm is employed to produce approximate samples from the posterior distribution given local feature estimates, derived from likelihood

Fig.2.MovesusedinMCMCsimulationontrees.iterative,EM-typealgorithmtomaximiselikelihood

(4)

(5)

(6)

wherethedataarewindowedwithacosinewindow,whosesizeistwicetheblockwidthatagivenscale,togivethedatausedintheestimator.Theindexdenotesiterationnumber;typically4-5iterationsaresuf- cienttogiveaccurateestimates.Figures3(b)-(c)showreconstructionsusingthe2DGaussiansineachblock(atcorrespondingblocksizes)basedontheMLfeatureesti-matesatblocksizesofandrespectively.Clearly,atlowerspatialresolutions,themodelcannoteas-ilydescribethepresenceofmultiplevesselswithinthewin-dow,suchasoccuratbifurcations,andtheresultinglow-amplitude,isotropicGaussiansarelocallythe‘best’de-scriptionoftheseregions.However,theseblockscanbemodelledaccuratelyathigherspatialresolutions.Thesec-ondsetofimagesshowshowthelocalestimatesfromdif-ferentwindowscalesisusedinacoarse- nestochasticsim-ulation,inordertogetaBayesianestimateoftheforest

structure.Afterthe rst200iterations,thescaleishalvedandtheappropriatelocalfeatureestimatesareusedtoguidethesampler;after3000,thescaleishalvedagain,asitisafter6000iterations,atwhichpoint,thehighestspatialres-olutionisreached.Thisapproachhasbeenfoundtospeedconvergencetotheequilibriumdistribution,whileavoidingbecomingtrappedinlocalmodes,inasimilarmannertomanycoarse- nealgorithms.Notethatoneiterationcon-sistsofthegenerationandacceptanceofasingleproposal(for‘editing’thetree).Inotherwords,100000iterationsiscomparable,intermsofcomputation,toasinglescanthroughtheimage.Ithasbeennotedfromexperimentsthatequilibriumisreachedinapproximately50000iterations,acomparativelylowburdencomputationally.

4.CONCLUSIONS

Someencouragingpreliminaryresultshavebeenachievedusingtheapproachdescribedinsection2,demonstratingitspotentialformodellingvascularstructuregloballyinacomputationallyef cientway.Fine-tuningthealgorithmwillleadtosigni cantimprovements.Thesewillinclude,forexample,theuseofthelocalestimatestoproduceinitialcon gurationsfortheMCMCalgorithm.Suchimprove-mentsarecurrentlybeingimplemented.Theworkisalsobeingtestedonothertypesofdataandextendedtothreedimensions.

5.REFERENCES

[1]P.DatlingerA.Pinz,S.BernoggerandA.Kruger.Map-pingthehumanretina.IEEETrans.MedicalImaging,17(1):606–619,1998.[2]S.P.Brooks.TheMarkovChainMonteCarloMethod

anditsApplication.TheStatistician,47:69–100,1998.[3]W.R.Gilks,S.Richardson,andD.J.Spiegelhalter.

MarkovChainMonteCarloinPractice.Chapman&Hall,1996.[4]M.E.Martinez-Perez,A.D.Hughes,A.V.Stanton,

A.S.Thom,A.A.Bharath,andK.H.Parker.Seg-mentationofretinalbloodvesselsbasedontheseconddirectionalderivativeandregiongrowing.InProc.ofIEEEICIP-99,pages173–176,Kobe,Japan,1999.[5]W.M.Wells,R.Kikinis,W.E.L.Grimson,and

R.Jolesz.AdaptivesegmentationofMRIdata.IEEETrans.MedicalImaging,15:429–442,1996.[6]D.L.WilsonandJ.A.Noble.Anadaptivesegmen-tationalgorithmforextractingarteriesandaneurysmsfromtime-of- ightMRAdata.IEEETrans.MedicalImaging,18(10):938–945,1999.

We describe a method for inferring tree-like vascular structures from 2D imagery. A Markov Chain Monte Carlo (MCMC) algorithm is employed to produce approximate samples from the posterior distribution given local feature estimates, derived from likelihood

(a)

(b)

(c)

Fig.3.(a)2Dretinalangiogramsize404404pixels.(b)Reconstructionofdatafrommodelparametersestimates

forblocksizesof64and(c)

16.

(a)

(b)

(c)

Fig.4.Estimatesfromthetree-basedsampler,at(a)200,(b)1000and(c)50000iterations,showinghowuseismadeofthelocalmultiresolutionfeatureestimates,inacoarse- neapproachtotheMCMCalgorithm.After50000iterations,fewchangesoccur.

本文来源:https://www.bwwdw.com/article/9kr4.html

微信扫码分享

《A Bayesian Approach to inferring vascular tree structure from 2D imagery.doc》
将本文的Word文档下载到电脑,方便收藏和打印
推荐度:
点击下载文档
下载全文
范文搜索
下载文档
Top